AiScribe

Train LLM from the coding session. Open source.

You code with AI. Ship 50 files. Next week you can't remember what you built or why. AiScribe fixes that.

It captures your git diffs and AI conversations. Structures them into summaries. Lets you export everything as training data to build your own model. Works with pi, Claude Code, Cursor, Codex, Aider, Windsurf, any AI coding tool.

How It Works

1. You code. AiScribe watches the git diff in the background.

2. Run aiscribe log. It sends your diff to an LLM and generates a summary. Use -c to capture the full AI conversation too.

3. Done. Markdown file saved in .aiscribe/sessions/. Git trackable. Greppable. Exportable.

What You Get Each Session

• Summary of what changed

• Files changed, insertions, deletions, risk levels

• Key decisions made during the session

• Suspicious changes flagged (unrelated files touched)

• Full chat log (optional, with -c)

Pick Your LLM

OpenRouter - 200+ models, one key Anthropic - Claude Sonnet, Opus OpenAI - GPT-4o, GPT-4o-mini DeepSeek - Cheap, fast Ollama - Local, free, no key needed

Configure once. Never asked again.

Export and Train

aiscribe export --format ai --output training-data.txt

Every session is training data. Fine tune a local model on your own coding style, decisions, and patterns.

Quick Start

npm install -g aiscribe cd your-project aiscribe log -c

GitHub · npm · Docs